Ground Vision: A Review of Machine Learning-Based Groundwater Level Prediction and Risk Analysis

Abstract

Groundwater is one of the most important sources of water for agriculture, domestic consumption, and industrial activities, yet its availability is strongly influenced by rainfall, recharge, temperature, evapotranspiration, hydrogeological characteristics, agricultural activity, irrigation demand, and human extraction. Conventional groundwater-prediction systems mainly forecast a future water-level value and give limited insight into the drivers behind that value or the effect of possible interventions. This paper reviews the application of Artificial Intelligence (AI), Machine Learning (ML), time-series forecasting, explainable AI, and decision-support techniques for groundwater-level prediction and management. The reviewed literature covers algorithms such as Random Forest, Support Vector Machine, XGBoost, Long Short-Term Memory networks, and hybrid ML–DL models, together with SHAP-based explainability, GIS-based risk mapping, drought early-warning systems, and IoT-based groundwater monitoring. Building on these approaches, GroundVision is proposed as a groundwater intelligence and policy-simulation platform that separates natural climate-driven groundwater behaviour from anthropogenic stress using a dual-layer prediction architecture, explains its predictions using SHAP, classifies locations into groundwater-risk categories, and allows administrators to test rainfall, irrigation, and drip-irrigation adoption scenarios through an interactive What-If Policy Sandbox. The review highlights that, while existing studies achieve strong prediction accuracy, they rarely combine forecasting, human-stress attribution, explainability, risk classification, and policy simulation within a single decision-support system, which is the gap that GroundVision aims to address.